feat: add executable editability and browser quality gates

[1beenlee]

Summary

• add tools/evaluate-edit.mjs with deterministic targeted, split, reorder, and compression before/after policies
• invoke the existing deck validator from the edit evaluator and block undeclared slide changes, :root design-token drift, navigation/script drift, and system-shell drift
• add positive/negative regression mutations over the unchanged fictional public example
• add tools/browser-qa.mjs that launches an already installed Chrome/Chromium and uses CDP through Node built-ins for real file:// runtime QA
• keep the browser sandbox enabled and route page network through a dead local proxy so QA does not give generated deck HTML an external-network path
• extend CI with syntax, editability regression, real browser QA, existing v0.3 Agent workflow regression, v0.2 generation harness regression, and public example validation
• document the boundary between mechanical containment/runtime evidence and semantic source/visual review

Scope guardrails

• examples/, templates/, prompts/, .agents/, .claude/: unchanged/read-only
• no npm/browser package install, Playwright/Puppeteer, model/provider API, external page fetch, hosted browser, deploy, Preview, artifact publishing, or user data
• no semantic/factual correctness claims from deterministic diff checks

Exact-head validation — PASS

GitHub Actions run 32744094707 on head 4361cafcfa988eabfd71990b63a89ad50daeb431 completed successfully.

node --check tools/prepare-deck.mjs                     # PASS
node --check tools/test-agent-workflow.mjs              # PASS
node --check tools/evaluate-edit.mjs                    # PASS
node --check tools/test-editability-eval.mjs            # PASS
node --check tools/browser-qa.mjs                       # PASS
node tools/test-editability-eval.mjs                    # PASS
node tools/browser-qa.mjs examples/hackathon-demo/index.html  # PASS
node tools/test-agent-workflow.mjs                      # PASS
node tools/test-generation-harness.mjs                  # PASS
node tools/validate-deck.mjs examples/hackathon-demo/index.html # PASS

Editability regression evidence

Positive/negative cases passed for:

• targeted edit accepted
• undeclared second-slide drift rejected
• :root design-token drift rejected
• navigation-script drift rejected
• declared split accepted
• unrelated split-side change rejected
• pure reorder accepted
• reorder content rewrite rejected
• constrained compression accepted
• undeclared removal rejected

Real Chrome runtime evidence

On installed Google Chrome in CI, the unchanged public example passed:

• file:// page attach/render, 10 slides
• generated TOC count/hash targets
• current / total page numbers
• no horizontal overflow at the tested 1440px viewport
• all slides within one tested viewport height
• initial active slide
• TOC click
• ArrowRight/ArrowLeft
• PageDown/PageUp
• End/Home
• Space
• direct hash navigation

Evidence boundary

This PR proves deterministic edit containment and the explicitly exercised browser-runtime behaviors. It does not claim semantic factual correctness, source fidelity of arbitrary edited prose, projector composition quality, 1280×800 behavior, or print/PDF quality; those remain separate review gates documented in docs/EDITABILITY_EVAL.md and docs/VALIDATION.md.